L4: Practical loss-based stepsize adaptation for deep learning
This work addresses the challenge of hyperparameter tuning in deep learning by providing a practical stepsize adaptation method that enhances optimizer performance, though it is incremental as it builds on existing optimizers.
The paper tackles the problem of stepsize adaptation in stochastic gradient descent by proposing a loss-based scheme that rescales gradients to achieve fixed predicted progress on the loss, resulting in consistent performance improvements for Adam and Momentum optimizers across various architectures and datasets without increased computational cost.
We propose a stepsize adaptation scheme for stochastic gradient descent. It operates directly with the loss function and rescales the gradient in order to make fixed predicted progress on the loss. We demonstrate its capabilities by conclusively improving the performance of Adam and Momentum optimizers. The enhanced optimizers with default hyperparameters consistently outperform their constant stepsize counterparts, even the best ones, without a measurable increase in computational cost. The performance is validated on multiple architectures including dense nets, CNNs, ResNets, and the recurrent Differential Neural Computer on classical datasets MNIST, fashion MNIST, CIFAR10 and others.